Global self-sustaining and local inheritance for source-free unsupervised domain adaptation

Abstract

In this paper, we investigate a practical problem called source-free unsupervised domain adaptation, which adapts a source-trained model to the target domain with unlabeled target data. To address this problem, we propose a novel GlObal self-sustAining and Local inheritance (GOAL) method. GOAL contains three components. (1) A backbone follows a mean teacher scheme. The teacher model serves as a smoothing functionality, facilitating a more consistent convergence of the student model. This capability alleviates the student model’s sensitivity to minor input data variations and enhances the overall robustness of the model. Additionally, disparities in predictions between the student and teacher models can be leveraged to identify potential noise in the data. (2) A Global Consistency Self-Sustaining mechanism for learning a stable, discriminative, and diverse prediction space. On the one hand, we employ neighbor samples and mean-teacher schemes to enhance the discriminability and stability of model predictions. On the other hand, non-neighbor samples are leveraged to augment the diversity of model predictions. Furthermore, to mitigate the impact of potential negative neighbors, we derive a weighting factor by incorporating both neighbor entropy and the top-nd similarity of features. (3) A Local Topology Inheritance mechanism to improve the semantic structure of the feature space. We construct a semantic topology graph based on the output predictions of the teacher model and subsequently transmit the teacher topology to the feature space of the student utilizing a local topology inheritance loss. Combining these three components, GOAL can effectively solve the source-free unsupervised domain adaptation. To the best of our knowledge, GOAL is the first attempt to perform topology inheritance for global consistency domain adaptation. Comprehensive experiments illustrate the effectiveness and superiority of GOAL in addressing source-free unsupervised domain adaptation.